Elastic Cloud Storage (ECS)

Elastic Cloud Storage (ECS) Overview and Architecture August 2017 A Dell EMC Technical Whitepaper

Revisions Date Description December 2015 Initial release ECS 2.2 May 2016 Updates for ECS 2.2.1 September 2016 Updates for ECS 3.0 August 2017 Updates for 3.1 The information in this publication is provided as is. Dell Inc. makes no representations or warranties of any kind with respect to the information in this publication, and specifically disclaims implied warranties of merchantability or fitness for a particular purpose. Use, copying, and distribution of any software described in this publication requires an applicable software license. Copyright August 2017 Dell Inc. or its subsidiaries. All Rights Reserved. Dell, EMC, and other trademarks are trademarks of Dell Inc. or its subsidiaries. Other trademarks may be the property of their respective owners. Published in the USA [8/23/2017] [Technical Whitepaper] [H14071] Dell believes the information in this document is accurate as of its publication date. The information is subject to change without notice. 2 Elastic Cloud Storage (ECS) H14071 1.10

Table of contents Revisions... 2 Executive Summary... 5 1 Introduction... 6 1.1 Audience... 6 1.2 Scope... 6 2 Value of ECS... 7 3 Architecture... 9 3.1 Overview... 9 3.2 ECS Portal and Provisioning Services... 10 3.2.1 ECS Management APIs... 11 3.2.2 ViPR SRM Support for ECS... 11 3.2.3 SNMP Support... 11 3.2.4 SYSLOG Support... 12 3.3 Data Services... 12 3.3.1 Object... 13 3.3.2 HDFS... 14 3.3.3 File... 15 3.4 Storage Engine... 15 3.4.1 Services... 15 3.4.2 Data and Metadata... 16 3.4.3 Data Management (Index and Partition Records)... 17 3.4.4 Data Flow... 19 3.4.5 Box Carting... 21 3.4.6 Space Reclamation... 22 3.5 Fabric... 22 3.5.1 Node Agent... 23 3.5.2 Life Cycle Manager... 23 3.5.3 Registry... 23 3.5.4 Event Library... 23 3.5.5 Hardware Manager (HWMgr)... 23 3.6 Infrastructure... 24 3.6.1 Docker... 24 3 Elastic Cloud Storage (ECS) H14071 1.10

3.6.2 Filesystem Layout... 25 3.7 Hardware... 26 3.7.1 ECS Appliance... 26 3.7.2 ECS Software on Industry Standard Hardware... 32 3.7.3 Network Separation... 33 4 Security... 36 4.1 Authentication... 36 4.2 Data Services Authentication... 36 4.3 Data At-Rest-Encryption (D@RE)... 37 4.3.1 Key Management... 37 5 Data Integrity and Protection... 39 5.1 Triple-mirrored... 39 5.2 Erasure Coding... 39 5.3 Checksums... 40 5.4 Compliance... 41 6 Deployment... 42 6.1 Single-site Deployment... 43 6.2 Multi-site Deployment... 44 6.2.1 Geo-Federation... 44 6.2.2 Geo-Replication... 44 6.2.3 Geo-Caching... 48 6.2.4 Temporary Site Outage... 48 6.3 Failure Tolerance... 50 7 Storage Efficiency... 53 8 Connectors and Gateways... 54 9 Conclusion... 55 A Resources... 56 A.1 References... 56 A.2 Support... 57 4 Elastic Cloud Storage (ECS) H14071 1.10

Executive Summary Dell EMC Elastic Cloud Storage (ECS ) is a software-defined, cloud-scale, object storage platform that combines the cost advantages of commodity infrastructure with the reliability, availability and serviceability of traditional arrays. With ECS, any organization can deliver scalable and simple public cloud services with the reliability and control of a private-cloud infrastructure. ECS provides comprehensive protocol support for unstructured (Object and File) workloads on a single, cloud-scale storage platform. With ECS, you can easily manage your globally distributed storage infrastructure under a single global namespace with anywhere access to content. ECS features a flexible software-defined architecture that is layered to promote limitless scalability. Each layer is completely abstracted and independently scalable with high availability and no single points of failure. Any organization can now realize true cloud-scale economics in their own data centers. 5 Elastic Cloud Storage (ECS) H14071 1.10

1 Introduction This document provides an in-depth architecture overview of Dell EMC Elastic Cloud Storage (ECS ), a turnkey software-defined cloud-scale object storage platform that combines the cost advantages of commodity infrastructure with the reliability, availability and serviceability of traditional arrays. 1.1 Audience 1.2 Scope This paper is intended for Dell EMC field personnel and customers who are interested in understanding the value and architecture of ECS. It provides an in-depth overview of ECS software, hardware, networking and services. It also provides links to other ECS information. This document focuses primarily on ECS architecture. It does not cover installation, administration, and upgrade procedures for ECS software or hardware. It also does not cover specifics on using and creating applications with the ECS APIs. It does provide an overview of available connectors and integrations and links to more information. Updates to this document are done periodically and coincides usually with a major release or new feature and functionality change. To get the latest version of this document, please download from this link. 6 Elastic Cloud Storage (ECS) H14071 1.10

2 Value of ECS ECS provides significant value for enterprises and service providers seeking a platform architected to support rapid data growth. The main advantages and features of ECS that enable enterprises to globally manage and store distributed content at scale include: Cloud Scale - ECS is an object storage platform for both traditional and next-gen workloads. It features a flexible software-defined architecture that promotes limitless scalability. Feature highlights: - Exabyte scale - Globally distributed object infrastructure - Supports billions of files of all types (big and small) in one storage platform - Easily expandable architecture Flexible Deployment - ECS has unmatched flexibility to deploy as an appliance, software-only solution, or in the cloud operated by Dell EMC or both. Features offered: - Appliance deployment (U Series, D series) - Software only deployment with support for certified or custom industry standard hardware - ECS Dedicated Cloud deployment that is fully hosted and managed by Dell EMC - Multi-protocol support: Object, File, Hadoop - Multiple workloads: IoT, Archive, Big data, Modern apps - Data domain and Isilon cloud tier: one destination for all primary storage Enterprise Grade - ECS provides customers more control of their data assets with enterprise class object, file and HDFS storage in a secure and compliant system with features such as: - Data at rest and replication across sites encryption - Reporting, policy based and event based record retention and platform hardening for SEC 17-A4 compliance including advanced retention management such as litigation hold and minmax governance - Authentication, authorization and access controls with Active directory/ldap - Integration with monitoring and alerting infrastructure (SNMP traps and SYSLOG) - Space reclamation - Enhanced enterprise capabilities (multi-tenancy, Swift multi-part upload, capacity monitoring, alerting, etc.) TCO Reduction - ECS can dramatically reduce TCO relative to traditional storage as well as public cloud storage. It even offers a lower TCO than Tape for LTR. Features include: - Global namespace - Small and large file performance - Seamless Centera migration - Raw capacity per rack increases from 3.8 PB to 6.2 PB - Lower total cost of ownership (TCO) than public cloud storage (TCS) - Lower TCO than other object storage solutions - Low management overhead - Small datacenter footprint - High storage utilization 7 Elastic Cloud Storage (ECS) H14071 1.10

The design of ECS is optimized for the following primary use cases: Geo Protected Archive ECS serves as a secure and affordable on-premise cloud for archival and long-term retention purposes. Using ECS as an archive tier can significantly reduce primary storage capacities. In ECS 2.2 and later, for cold archives, a 10+2 erasure coding scheme is available in which a chunk is broken down into 10 data fragments and 2 coding (parity) fragments. This allows for better storage efficiencies for this particular use case. Global Content Repository Unstructured content repositories containing images, videos etc. are currently stored in high cost storage systems making it impossible for businesses to cost-effectively manage massive data growth. ECS enables any organization to consolidate multiple storage systems into a single, globally accessible and efficient content repository. Storage for Internet of Things The Internet of Things offers a new revenue opportunity for businesses who can extract value from customer data. ECS offers an efficient IoT architecture for unstructured data collection at massive scale. With no limits on the number of objects, the size of objects, or metadata, ECS is the ideal platform to store IoT data. ECS also streamlines analytics because the data can be analyzed directly on the ECS platform without requiring time consuming ETL (Extract, transform, load) processes. Just point the Hadoop cluster at ECS to start running queries. As an object storage system, ECS can store all the IoT data in a very cost effective manner and leverage its built-in data checking to ensure data integrity. Video Surveillance Evidence Repository In contrast to IoT data, video surveillance data has a much smaller object storage count, but a much higher capacity footprint per file. While data authenticity is important, data retention is not as critical. ECS can be a low-cost landing area or secondary storage location for this data. Video management software can leverage ECS rich metadata to tag files with important details like camera location, retention requirement and data protection requirement. Also, ECS metadata can be used to set the file to a read-only status to ensure a chain of custody on the file. Modern Applications ECS provides a single turnkey platform designed for modern application development, management and analytics. ECS is made to support next-gen web, mobile and cloud applications. Multi-site read/writes with strong consistency make developers job much easier. As the ECS capacity changes and grows, developers never need to recode their apps. Data Lake For organizations of any size, ECS establishes a data lake foundation. It fully maximizes user data with its powerful HDFS service, making Big Data applications and analytics a production reality. It features In-place analytics capabilities to reduce risk, resources and time-to-results. Cloud Backup ECS can be used as a cloud target backup for customer s primary data. For instance, utilizing CloudPools to tier data from Isilon to ECS. Third party cloud backup solutions can also typically be redirected to ECS as the cloud backup target. 8 Elastic Cloud Storage (ECS) H14071 1.10

3 Architecture ECS is architected with certain design principles, such as global namespace with strong consistency; scaleout capability, secure multi-tenancy; and superior performance for both small and large objects. ECS was built as a completely distributed system following the principle of cloud applications. The ECS software running on commodity nodes forms the underlying cloud storage, providing protection, geo replication, and data access. This section will go in-depth into the ECS architecture and design of both the software and hardware. 3.1 Overview ECS provides a software-defined cloud storage platform that can be deployed on a set of qualified industry standard hardware or a turnkey storage appliance. At a high level ECS is composed of the following main components: ECS Portal and Provisioning Services provides a Web-based portal that allows self-service, automation, reporting and management of ECS nodes. It also handles licensing, authentication, multi-tenancy, and provisioning services. Data Services provides services, tools and APIs to support Object, and HDFS and NFSv3. Storage Engine responsible for storing and retrieving data, managing transactions, and protecting and replicating data. Fabric provides clustering, health, software and configuration management as well as upgrade capabilities and alerting. Infrastructure uses SUSE Linux Enterprise Server 12 as the base operating system for the turnkey appliance or qualified Linux operating systems for industry standard hardware configuration. Hardware offers a turnkey appliance, qualified industry standard hardware, or hosted hardware. Figure 1 shows a graphical view of these layers. Each layer is described in more detail in the subsections that follow. Figure 1 - ECS Architecture Overview 9 Elastic Cloud Storage (ECS) H14071 1.10

3.2 ECS Portal and Provisioning Services ECS management is done through the ECS Portal and provisioning services. ECS provides a Web-based GUI that allows you to manage, license, and provision ECS nodes. The portal has comprehensive reporting capabilities that include: capacity utilization per site, storage pool, node and disk performance monitoring on latency, throughput, transactions per second, and replication progress and rate diagnostic information, such as node and disk recovery status and per-node statistics on hardware and process health, which helps identify performance and system bottlenecks. The ECS dashboard provides overall system-level health and performance information. This unified view enhances overall system visibility. Alerts notify users about critical events, such as capacity limits, quota limits, disk or node failures, or software failures. ECS also provides a command-line interface to install, upgrade, and monitor ECS. Access to nodes for command-line usage is done via SSH. A screenshot of the ECS Dashboard appears in Figure 2. Figure 2 - ECS Dashboard. 10 Elastic Cloud Storage (ECS) H14071 1.10

3.2.1 ECS Management APIs You can also manage ECS using RESTful APIs. The management APIs allows users to administer the ECS system within their own tools, scripts or existing applications. It can also be used to create new monitoring applications. The ECS Portal and command-line tools were built using the ECS REST Management APIs. 3.2.2 ViPR SRM Support for ECS In ECS 2.0 and later, VIPR SRM incorporates ECS monitoring information and provides dashboards and reports relating to object utilization. For instance, ViPR SRM object dashboards now include a summary of configured usable capacity, capacity trends, and configurable usable capacity by service level. The inventory namespace report provides detailed information on quota status and used capacity by namespace. Namespace chargeback reports show total used capacity for local and remote systems and the total number of objects per namespace to identify service levels, cost contributors and charges. Bucket-level reports provide details on the number of objects as well as used capacity and the percentage of quota used by a bucket. Through ViPR SRM you also can view performance and capacity trends over a specified period of time. Please refer to ViPR SRM on Dell EMC s website for more information. Figure 3 shows a screenshot of a Namespace Chargeback Report in ViPR SRM. Figure 3 - Namespace Chargeback Reports Namespace chargeback reports identify service levels, cost contributors and charges 3.2.3 SNMP Support Large enterprises, such as financial companies, rely heavily on Simple Network Management Protocol (SNMP) for collecting information related to devices managed on networks. In ECS 2.2.1 and later, initial support for SNMP allows for basic queries. There is an SNMP agent (snmpd) enabled and set to start at the operating system level to support Management Information Base (MIB) which is a set of network objects that can be managed or monitored using SNMP. Current support provides basic queries of node level statistics which include CPU and memory utilization and number of processes. All nodes can be monitored; however, each node would need to be queried independently. An example of an SNMP query for the number of processes: 11 Elastic Cloud Storage (ECS) H14071 1.10

# snmpget -v 2c -c public 10.249.251.176 iso.3.6.1.2.1.25.1.6.0 iso.3.6.1.2.1.25.1.6.0 = Gauge32: 41 To enable and disable this feature, use setrackinfo command. Refer to the ECS Installation Guides in SolVe Desktop (Procedure Generator) for more information on configuration of SNMP. In addition to SNMP basic queries, SNMP traps is an optional feature available in ECS 3.0 or later to notify and alert customer s network management system when ECS critical errors and warnings occur. SNMP traps provide better visibility of the state of ECS and overall improved product serviceability. Up to 10 SNMP trap destination servers can be configured. Also, it supports SNMP v2 and SNMP v3 (User-based Security Model). For SNMP v3 authentication, the following are supported: Authentication protocols - MD5 and SHA Privacy protocols - DES, AES128, AES192, and AES256 Security Levels noauthnopriv, authnopriv and authpriv Examples of critical alerts that generate traps include disk failures, node failures, fabric agent failure, or network interface down. There are also informational and warning events that can generate traps such as license expiration, namespace and bucket hard quota limits has exceeded, node or disk is in suspect mode, or network interface IP address has been updated. For complete list of SNMP traps and how to configure, refer to the ECS System Administrators Guide. 3.2.4 SYSLOG Support Introduced in ECS 3.0 is syslog support to forward ECS event information to remote centralized log server(s). The ECS logs that currently can be forwarded include: ECS Audit Logs and Alerts Operating System Logs It supports both UDP and TCP based communication with syslog servers. It also is able to forward to multiple redundant syslog servers that are active. Configuration of syslog such as addition, deletion and, modification of syslog servers and specification of severity threshold of logs to be forwarded is done via the portal or ECS REST Management APIs. Only the system administrator has the privilege to perform syslog management operations. ECS syslog support is a distributed service and resilient to node failures. For more information on how to configure, refer to the ECS System Administrators Guide. 3.3 Data Services Access to data stored in ECS is through Object, HDFS and NFS v3 protocols. In general, ECS provides multiprotocol access, meaning data ingested through one protocol can be accessed through another. For example, data can be ingested through S3 and modified through NFSv3 or HDFS (or vice versa). There are some exceptions to this multi-protocol access due to protocol semantics and representations of how the protocol was designed. Table 1 highlight the Object APIs and protocol supported and which protocols interoperate. 12 Elastic Cloud Storage (ECS) H14071 1.10

Table 1 - ECS Supported Data Services Protocols Supported Interoperability S3 Additional capabilities like Byte Range HDFS, NFS, Swift Object Updates and Rich ACLS Atmos NFS (path-based objects only and not Version 2.0 object ID style based) Swift V2 APIs and Swift and Keystone v3 HDFS, NFS, S3 Authentication CAS SDK v3.1.544 or later N/A HDFS Hadoop 2.7 compatibility S3, NFS, Swift NFS NFSv3 S3, Swift, HDFS, Atmos (path-based objects only and not object ID style based) The data services, which are also referred to as head services, are responsible for taking client requests, extracting required information, and passing it to the storage engine for further processing (e.g. read, write, etc.). In 2.2H1 and later, all head services had been combined to one process running on the infrastructure layer to handle each of the protocols called dataheadsvc, in order to reduce overall memory consumption. This process is further encapsulated within a Docker container named object-main, which runs on every node within the ECS system. The Infrastructure section of this document covers this topic in more detail. Also, in order to access objects via the above protocols, certain firewall ports need to be opened. For more information on ports refer to the ECS Security Configuration Guide. 3.3.1 Object For object access, ECS provides industry-standard object APIs. ECS supports S3, Atmos, Swift and CAS APIs. With the exception of CAS, objects or data are written, retrieved, updated, and deleted via HTTP or HTTPS calls of GET, POST, PUT, DELETE and HEAD. For CAS, standard TCP communications and specific access methods and calls are used. In addition, ECS provides a facility for metadata search of objects. This enables ECS to maintain an index of the objects in a bucket, based on their associated metadata, allowing S3 object clients to search for objects within buckets based on the indexed metadata using a rich query language. Search indexes can be up to 30 metadata fields per bucket and are configured at the time of bucket creation through the ECS Portal, ECS Management REST API, or S3 REST API. In ECS version 3.1 or later, metadata search feature can be enabled on buckets with server side encryption enabled; however, any indexed user metadata attribute utilized as a search key will not be encrypted. For CAS objects, CAS query API provides similar ability to search for objects based on metadata that is maintained for CAS objects, and does not need to be enabled explicitly. For more information on ECS APIs and APIs for metadata search, see the ECS Data Access Guide. For Atmos and S3 SDKs, refer to the GitHub site: Dell EMC Data Services SDK. For CAS, refer to the Centera Community site. You can also access numerous examples, resources and assistance for developers in the ECS Community. 13 Elastic Cloud Storage (ECS) H14071 1.10

If you have access to an ECS system, there are existing client applications that provide a way to quickly test or access data stored in ECS for example, the S3 Browser and Cyberduck. Another option is to use the ECS Test Drive, which allows you to gain access to an ECS system for testing and development purposes. After registering for ECS Test Drive, you receive REST endpoints and can create and/or manage user credentials for each of the object protocols. With the endpoints, you can use clients like S3 Browser and Cyberduck to read and write objects. 3.3.2 HDFS ECS can store Hadoop file system data, which allows organizations to create Big Data repositories on ECS that Hadoop analytics can consume and process. The HDFS data service is compatible with Apache Hadoop 2.7, with support for fine-grained ACLs and extended filesystem attributes. In ECS 2.2 and later, ECS has been integrated with Ambari, which allows you to easily deploy ECS HDFS clients (jar file) and specify ECS HDFS as the default filesystem in a Hadoop cluster. As illustrated in Figure 4, an ECS HDFS Client file is installed on each node within a Hadoop cluster. ECS provides file system and storage functionality equivalent to what name nodes and data nodes do in a Hadoop deployment. ECS streamlines the workflow of Hadoop with direct attached storage (DAS) by eliminating the need for migration of data to a local Hadoop DAS and/or creating a minimum of three copies. Figure 4 - ECS HDFS Other enhancements added in ECS 2.2 for HDFS include the following: Proxy User Authentication impersonation for Hive, HBase, and Oozie. Security - server-side ACL enforcement and addition of Hadoop superuser and superuser group as well as default group on buckets. 14 Elastic Cloud Storage (ECS) H14071 1.10

ECS 2.2 has been validated and tested with Hortonworks (HDP 2.5.3). It also has support for services such as YARN, MapReduce, Pig, Hive/Hiveserver2, HBase, Zookeeper, Flume, Spark and Sqoop. For more information, refer to Enabling Hadoop with Dell EMC Elastic Cloud Storage whitepaper. 3.3.3 File ECS 2.2.1 and later includes native file support with NFSv3. The main features for the NFSv3 file data service include: Global namespace ability to access the file from any node at any site. Global locking ability to lock files from any node at any site (shared and exclusive locks, range based locks, and mandatory locks) Multi-protocol access ability to access data created by object (S3, Swift, Atmos), HDFS, and NFS. NFSv3 is configurable via the ECS Portal. NFS exports, permissions, and user group mappings are created through the API and/or the portal. In addition, NFS clients such as Solaris, Linux, and Windows can mount the export specified in the ECS Portal using namespace and bucket names. For example: mount t nfs o vers=3 <ip of node or DNS name>:<export path( i.e. /namespace/bucket)>. To achieve client transparency during a node failure, a load balancer is recommended. ECS has tightly integrated the other NFS server implementations, such as lockmgr, statd, nfsd, and mountd, hence, these services are not dependent on the infrastructure layer (host operating system) to manage. NFS v3 support has the following features: No design limits on the number of files or directories File write size can be up to 4TB Ability to scale across 8 sites with a single global namespace/share Support for Kerberos and AUTH_SYS Authentication NFS file services process NFS requests coming from clients; however, data is stored as objects, similar to the object data service. An NFS file handle is mapped to an object id. Since the file is basically mapped to an object, NFS has features similar to the object data service, including: Quota Management at the bucket level Encryption at the object level 3.4 Storage Engine At the core of ECS is the storage engine. This layer contains the main components that are responsible for processing requests as well as storing, retrieving, protecting and replicating data. This section describes the design principles and the way that data is represented and handled internally. Data flow for reads and writes is also described. 3.4.1 Services ECS has a layered architecture, with every function in the system built as an independent layer. This design principle makes each layer horizontally scalable across all nodes in the system and ensures high availability. 15 Elastic Cloud Storage (ECS) H14071 1.10

The ECS storage engine includes the layers shown in Figure 5, which run on top of the infrastructure and hardware component. Figure 5 - Storage Engine Layers The services of the Storage Engine are encapsulated within a Docker container and installed on each ECS node, providing a distributed and shared service. 3.4.2 Data and Metadata The details of data stored in ECS can be summarized as follows: Data the actual content that needs to be stored (image, document, text, etc.) System Metadata information and attributes relating to the data. System metadata can be categorized as follows: - Identifiers and descriptors A set of attributes used internally to identify objects and their versions. Identifiers are either numeric ids or hash values which are not of use outside the ECS software context. Descriptors define information such as type of content and encoding. - Encryption keys in encrypted format Object encryption keys are included as part of system metadata after being encrypted by KEKs(key encryption keys) - Internal flags A set of flags to track if byte range updates or encryption are enabled, as well as to coordinate object locks, caching and deletion - Location information A set of attributes with index and data location information such as byte offsets. - Timestamps A set of attributes that track time of object creation, modification and expiration. - Configuration/tenancy information Object names, namespace name, zone/vdc name and access control information for objects. 16 Elastic Cloud Storage (ECS) H14071 1.10

Custom User Metadata user-defined metadata, which provides further information or categorization of the data that is being stored. Custom metadata is formatted as key-value pairs that are sent with a write request (e.g. Client=Dell EMC, Event=DellEMCWorld, ID=123). All types of data, including system and custom metadata, are stored in chunks. A chunk is a 128MB logical container of contiguous space. Note that each chunk can have data from different objects, as shown in Figure 6. ECS uses indexing to keep track of all the parts of an object that may be spread across different chunks and nodes. Chunks are written in an append-only pattern, meaning, an application cannot modify/delete existing data within a chunk but rather updated data is written in a new chunk. Therefore, no locking is required for I/O and no cache invalidation is required. The append-only design also simplifies data versioning. Old versions of the data are maintained in previous chunks. If S3 versioning is enabled and an older version of the data is needed, it can be retrieved or restored to a previous version using the S3 REST API. Figure 6 - Chunk (can store data from different objects) All chunks are triple-mirrored to multiple nodes to protect data against drive or node failures. When data-only chunks fill up to 128MB and/or are sealed, the chunk is erasure coded for storage efficiency and spread across multiple nodes for data protection. Once a chunk has been erasure coded and protected across the nodes and disks, the mirrored copies are discarded. The chunks containing index and metadata are triplemirrored and are not erasure coded. A more detailed description of triple mirroring and erasure coding are provided in the Data Protection section of this document. 3.4.3 Data Management (Index and Partition Records) ECS uses logical tables to keep track of where data and metadata chunks are stored on disk. The tables hold key-value pairs to store information relating to the objects. A hash function is used to do fast lookups of values associated with a key. These key-value pairs are eventually stored in a B+ tree for fast indexing of data locations. By storing the key-value pair in a balanced, searched tree like a B+ Tree, the location of the data and metadata can be accessed quickly. In addition, to further enhance query performance of these logical tables, ECS implements a two-level log-structured merge (LSM) tree. Thus there are two tree-like structures where a smaller tree is in memory (memory table) and the main B+ tree resides on disk. So lookup of key-value pairs will be first looked up in memory and if value is not memory, it will look at the main B+ tree on disk. Entries in these logical tables are first recorded in journal logs and these logs are written to disks as chunks and triple-mirrored. The journals keep track of index transactions not yet committed to the B+ tree. After the transaction is logged into a journal, the memory table is updated. Once the table in the memory becomes full or after a certain period of time, the table is eventually merged sorted or dumped to B+ tree on disk. Figure 7 illustrates this process. 17 Elastic Cloud Storage (ECS) H14071 1.10

Figure 7 - Logical Tables One example of a logical table is the object table (Figure 8) which contains the name of an object and the chunk location (Chunk 1) at a certain offset and length within that chunk. In this table, the object name is the key to the index and the value is the chunk location. The index layer within the Storage Engine is responsible for the object-name-to-chunk mapping. Figure 8 Object Table Object Name Chunk Location ImgA FileB C1:offset:length C2:offset:length C3:offset:length Another table referred to as the Chunk Table records which node the chunk resides on. For data integrity, the chunks are triple-mirrored to different disks on different nodes and this table keeps track of where the chunk location resides. This table keeps track of which nodes the chunk resides on, the disk within the node, the file within the disk, the offset within that file and the length of the data, as shown in Figure 9. Figure 9 Chunk Table Chunk ID C1 Location Node1:Disk1:File1:Offset1:Length Node2:Disk2:File1:Offset2:Length Node3:Disk2:File6:Offset:Length 18 Elastic Cloud Storage (ECS) H14071 1.10

As mentioned previously, ECS was designed to be a distributed system such that storage and access of data are spread across the nodes. Since these tables can get very large, these logical tables are divided into partitions and assigned to different nodes. The node then becomes the owner of that partition or section of the table. So to get the location of a chunk would require retrieving the information from the partition owner. A partition record table is illustrated in Figure 10 below. Figure 10 - The Partition Records Table Partition ID Owner P1 Node 1 P2 Node 2 P3 Node 3 If the node goes down, another node takes ownership of the partition. The logical tables owned by the unavailable node get recreated on a new node and that node becomes the partition owner. The table is recreated by using the B+ tree root stored on disk and replaying the journals also stored on disk. As previously mentioned the B+ tree and journals are stored in chunks (just like data) and are triple-mirrored. Figure 11 shows the failover of partition ownership. Figure 11 - Failover of Partition Ownership 3.4.4 Data Flow Data can be accessed by any of the ECS nodes and written to three different nodes (triple-mirrored) within ECS for data protection and resiliency. Also, prior to writing data to disk, ECS runs a checksum function and stores the result; then data is evaluated for compression. If in the first few bytes of data is compressible it will compress the data, otherwise it will not for performance. Similarly with reads, data is decompressed and checksum is validated. The data flow of reads and writes are described below. 19 Elastic Cloud Storage (ECS) H14071 1.10

The data flow for writes is as follows (shown in Figure 12): 1. A create object is requested from client and is directed to one of the nodes (Node 1 in this case) who will process the request. 2. The data are written to a new chunk or appended to an existing chunk and triple-mirrored to three different nodes in parallel (initiated by Node 1 in this example). 3. Once the three copies are acknowledged to have been written, an update to the logical tables (index) of partition owner occurs with name and chunk location. 4. The partition owner of index ( Node 3 in this case) records this write transaction into the journal logs which is also written into a chunk on disk and triple mirrored to three different nodes in parallel (initiated by Node 3). 5. Once the transaction has been recorded in the journal logs, an acknowledgement is sent to the client. In a background process, journal entries are processed and persisted onto disk in B+ trees as chunks and triple-mirrored. Figure 12 - Write Data Flow The data flow for reads include the following (shown in Figure 13): 1. An object read request is received and directed to any available node to process. Node 1 in this example. 2. As mentioned in the Data Management section of this paper, there are logical tables (key-value pairs) containing information to the location of the object, like object table and chunk table. These tables are partitioned and are distributed among the nodes to own and manage. So, object request will be processed by Node 1 in this case. Node 1 will utilize a hash function using the object name ID to determine which node is the partition owner of the logical table where this object information resides. In this example, Node 2 is owner and thus Node 2 will do a lookup in the logical tables to get location of chunk. In some cases, the lookup can occur on two different nodes, for instance when the location is not cached in logical tables of Node 2. 20 Elastic Cloud Storage (ECS) H14071 1.10

3. From the previous step, location of chunk is provided to Node 1 who will then issue a byte offset read request to the node that holds the data, Node 3 in this example, and will send data to Node 1. 4. Node 1 will send data to requesting client. Figure 13 Read Data Flow 3.4.5 Box Carting ECS has a unique and inherent feature called box-carting which helps performance for data writes. Boxcarting aggregates multiple small data objects queued in memory and then write them in a single disk operation, up to 2MB of data. This improves performance by reducing the number of roundtrips to process individual writes to storage. Figure 14 provides an illustration of box-carting. Figure 14 - Box-Carting 21 Elastic Cloud Storage (ECS) H14071 1.10

As for writes of large objects, all nodes within ECS can process write requests for the same object simultaneously and write to different set of disks, taking advantage of all the spindles in the ECS cluster. Thus, ECS can ingest and store small and large objects efficiently. 3.4.6 Space Reclamation ECS writes data to chunks and, during updates and deletes, there will be blocks of data that are no longer referenced and used. In ECS 3.0, space reclamation or garbage collection has been enhanced to not only reclaim space when a chunk is completely empty, but also reclaim space when a chunk is partially empty. There are three approaches to space reclamation supported in ECS 3.0 and later releases: Normal Garbage Collection when entire chunk is garbage, reclaim space. Partial Garbage Collection by Merge when chunk is 2/3 garbage, reclaim the chunk by merging the valid parts of chunk with other partially filled chunks to a new chunk and reclaim the space. Partial Garbage Collection with Compaction when chunk is greater than 1/3 and less than 2/3 garbage, the valid parts of the chunk are re-erasure coded to protect the valid data. This involves treating the garbage ranges within the chunk as zeros for re-calculating the erasure code of the valid data. Once the valid data has been re-erasure coded the garbage ranges are then eligible to be reclaimed. If the chunk is replicated to another site within a replication group, utilize a mathematically XOR approach to replace the garbage space at remote sites before the garbage chunk or chunk ranges are freed for re-use. This method requires an XOR encode of a garbage chunk or a set of partial chunks whose garbage sum is equal to or greater than 128MB with a new chunk that is being shipped to same site where the garbage ranges were replicated to; and then, XOR chunk again at remote site to replace the garbage ranges of the chunk copies with the new chunk. Once this operation has been done, the garbage ranges are eligible to be reclaimed. This avoids the extra WAN traffic specifically when there are 3 or more sites. With the support of partial chunk reclamation, ECS provides better overall space utilization. Note that the XOR process done in partial garbage collection should not be confused with the XOR approach used for storage efficiency in three or more sites. For detailed description of the new space reclamation introduced in ECS 3.0, please refer to the ECS Garbage Collection Process Overview whitepaper, internal only. In ECS 3.1 or later, normal garbage collection and partial garbage collection by merge will be enabled by default. 3.5 Fabric The Fabric layer provides clustering, system health, software management, configuration management, upgrade capabilities, and alerting. It is responsible for keeping the services running and managing resources such as disks, containers, the firewall, and the network. It tracks and reacts to environment changes such as failure detection and provides alerts related to system health. The fabric layer has the following components to manage the overall system: Node agent manages host resources (disks, the network, containers, etc.) and system processes. Lifecyle Manager application lifecycle management which involves starting services, recovery, notification, and failure detection. Persistence Manager provides the coordination and synchronization of the ECS distributed environment. 22 Elastic Cloud Storage (ECS) H14071 1.10

Registry stores all the Docker images for ECS. Event Library holds the set of events occurring on the system. Hardware Manager (HWMgr) provides status, event information and provisioning of the hardware layer to higher level services. These services have been integrated to the Fabric Agent to support industry standard hardware. 3.5.1 Node Agent The node agent is a lightweight agent written in Java that runs natively on every ECS node. Its main duties include managing and controlling host resources (Docker containers, disks, the firewall, the network), and monitoring system processes. Examples of management include formatting and mounting disks, opening required ports, ensuring all processes are running, and determining public and private network interfaces. It also has an event stream that provides ordered events to a lifecycle manager to indicate events occurring on the system. A Fabric CLI allows you to diagnose issues and look at overall system state. 3.5.2 Life Cycle Manager The lifecycle manager runs on a subset of nodes (three to five instances) and manages the lifecycle of applications (object-main) running on various nodes. Each lifecycle manager is responsible for tracking a number of nodes. Its main goal is to manage the entire lifecycle of the ECS application from boot to deployment, including failure detection, recovery, notification and migration. It looks at the node agent streams and drives the agent to handle the situation. When a node is down, it responds to failures or inconsistencies in the state of the node by restoring the system to a known good state. If a lifecycle manager instance is down, another one takes its place if available. 3.5.3 Registry The registry contains all the ECS Docker images used during installation, upgrade and node replacement. A Docker container called fabric-registry runs on one node within the ECS rack and holds the repository of ECS Docker images and information required for installations and upgrades. Although the registry is available on one node, all Docker images are locally cached on all nodes. 3.5.4 Event Library The event library is used within the Fabric layer to expose the lifecycle and node agent event streams. Events generated by the system are persisted onto shared memory and disk to provide historical information on the state and health of the ECS system. These ordered event streams can be used to restore the system to a specific state by replaying the ordered events stored. Some examples of events include node events such as started, stopped or degraded. 3.5.5 Hardware Manager (HWMgr) The hardware manager has been integrated to the Fabric Agent in order to support industry standard hardware. Its main purpose is to provide hardware specific status and event information, and provisioning of the hardware layer to higher level services within ECS. 23 Elastic Cloud Storage (ECS) H14071 1.10

3.6 Infrastructure Each ECS node runs a specific operating system. ECS Appliance currently runs SuSE Linux Enterprise 12 which acts as the ECS infrastructure. For the ECS software deployed on custom industry standard hardware the operating system can be RedHat Enterprise Linux (RHEL) or CoreOS. Custom deployments are done via a formal request and validation process. Docker is also installed on the infrastructure to deploy the encapsulated ECS layers described in the previous sections. Because ECS software is written in Java, the Java Virtual Machine (JVM) is installed as part of the infrastructure. 3.6.1 Docker ECS runs on top of the operating system as a Java application and is encapsulated within several Docker containers. The containers are isolated but share the underlying operating system resources and hardware. Some parts of ECS software run on all nodes and some run on one or some nodes. The components running within a Docker container include: object-main contains the resources and processes relating to the data services, storage engine, portal and provisioning services. Runs on every node in ECS. fabric-lifecycle contains the processes, information and resources required for system-level monitoring, configuration management and health management. Depending on the number of nodes in the system, there will be an odd number of fabric-lifecycle instances running. For example, there will be three instances running on a four-node system and five instances for an eight-node system. fabric-zookeeper centralized service for coordinating and synchronizing distributed processes, configuration information, groups and naming services. It is referred to as the persistence manager and runs on odd number of nodes, for instance, three for a four-node system and five for an eight node system. fabric-registry location or registry of the ECS Docker images. Only one instance runs per ECS rack. There are other processes and tools that run outside of a Docker container namely the Fabric node agent and hardware abstraction layer tools. Figure 15 below provides an example of how ECS is run on a four node deployment. 24 Elastic Cloud Storage (ECS) H14071 1.10

Figure 15 - Docker Containers and Agents on Eight Node Deployment Example Figure 16 shows a view of the object-main container in one of the nodes of an ECS system. The listing provides the some of the services running in the Docker container. Figure 16 - Processes, resources, tools and binaries in object-main container 3.6.2 Filesystem Layout ECS Each ECS node is directly connected to a set of commodity disks. Each disk is formatted and mounted as an XFS filesystem and has a unique identifier (UUID). Data are stored in chunks and the chunks are stored in files within the filesystem. Each disk partition or filesystem is filled with files that are 10GB in size. These files are created during installation such that contiguous blocks are allocated for the files. The number of files within each disk depends on the disk size. For instance, a 1TB disk will have 100, 10GB files. The names of 25 Elastic Cloud Storage (ECS) H14071 1.10

the files inside the disk are 0000, 0001, 0002, and so forth. Figure 17 provides a screenshot of the XFS filesystems and 10GB files populated on one node of the ECS system. Figure 17- Mounted XFS and 10G Files Storing the Chunks 3.7 Hardware ECS software can run on a turn-key Dell EMC appliance or on industry standard hardware. Flexible entry points exist with the ability to rapidly scale to petabytes and exabytes of data. With minimal business impact, you can scale ECS linearly in both capacity and performance by simply adding nodes and disks to your environment. The basic hardware required to run ECS includes a minimum of four compute nodes with data disks, a pair of 10 GbE switches for data, and a 1 GbE management switch. 3.7.1 ECS Appliance The ECS appliance currently supports two generations of the U-Series model and a denser model, D-Series. This section will only highlight the hardware components required to run ECS software. For additional documentation on installing and cabling ECS hardware see the following documents: Dell EMC ECS Product Documentation Index - ECS Planning Guide - ECS Hardware and Cabling Guide - ECS System Administrator s Guide SolVe Desktop (Procedure Generator) which has information on Switch Configuration Files and Installing ECS Software Guide. This link may require access to the Dell EMC support website. Figures 18, 19, and 20 in the next sections provide a view of the current U-Series (Gen-1 and Gen-2) and D- Series models. Please refer to the ECS Appliance Specification Sheet for more information on ECS models. The data sheets also contain information about power requirements for each model. 26 Elastic Cloud Storage (ECS) H14071 1.10

3.7.1.1 U- Series A standard U-Series appliance contains two 10 GbE switches, one 1 GbE switch, and four or eight x86 server nodes and disk array enclosures (DAE). Each DAE is connected to one x86 node by SAS. For instance, a five-node model has five disk array enclosures and an eight-node model has 8 disk enclosures. For the U- Series, a DAE can have up to 60 disks. Figure 18 - ECS Appliance Gen-1 Models Figure 19 - ECS Appliance Gen-2 Models 27 Elastic Cloud Storage (ECS) H14071 1.10

3.7.1.2 D-Series The D-series are the denser models. They are targeted for very large deployment and as a replacement for tape archives. The D-series offer between 4.5 PB and 7.8 PB of max raw capacity options. It is also designed for energy efficiency. Configurations of the D-Series have a minimum of eight x86 nodes and can hold up to 98 disks. It also has two 10 GbE switches and one 1 GbE switch. These configurations offer a lower TCO comparable to tape. Minimum ECS 2.2.1 release is required for D-Series. Figure 20- ECS Appliance D-Series Model 3.7.1.3 ECS Appliance Nodes ECS can come with 1 to 4 nodes in a 2U chassis as shown in Figure 21. Server specifications are as follows: 4 nodes in 2U with two CPUs per node - Gen 1-2.4 GHz four-core Ivy Bridge - Gen 2 and D-series 2.4 GHz six-core Haswell CPUs 64GB memory per node Dual 10Gb Ethernet (NIC-bonding makes them appear as a single port) Dual 1Gb Ethernet Dedicated 1Gb Ethernet remote diagnostics 4 x 2.5 drive slots per node Single (1) 400GB SSD system disk Single DOM for OS boot Two SAS interfaces per node 28 Elastic Cloud Storage (ECS) H14071 1.10

Figure 21 Example of a 4 Node (Front and Rear View) 3.7.1.4 10 GbE Switches Data Two 10 GbE, 24-port or 52-port Arista switches are used for data transfer to and from customer applications as well as for internal node-to-node communications. These switches are connected to the ECS nodes in the same rack. The switches employ the Multi-Chassis Link Aggregation (MLAG) feature, which logically links the switches and enables active-active paths between the nodes and customer applications. This configuration results in higher bandwidth while preserving resiliency and redundancy in the data path. Any networking device supporting static LAG or IEEE 802.3ad LACP can connect to this MLAG switch pair. Finally, because the switches are configured as MLAG, these two switches appear and act as one large switch. Figure 22 displays an example of the front view of these 2 switches. Figure 22 An Example of 10GbE Arista Switches (Front View) 3.7.1.5 1 GbE Switch - Management The 52-port 1 GbE Arista switch is used by ECS for node management and out-of-band management communication between the customer s network and the Remote Management Module (RMM) ports of the individual nodes. The main purpose of this switch is for remote management and console, install manager (PXE booting), and enables rack management and cluster wide management and provisioning. Figure 23 shows a front view of an Arista 1 GbE management switch. Figure 23 Example of 1 GbE Arista Switch (Front View) In addition to Arista, there is now support for Cisco 52 port 1 GbE switch for management. This switch is meant to support customers who have strict Cisco only requirements. It is available only for new racks and is not supported to replace Arista management switches in existing racks. Configuration file will be pre-loaded in manufacturing and will still be under control of Dell EMC personnel. ECS 3.0 (with patches) is the minimum to support the Cisco management switch. Figure 24 shows front view of a Cisco 1 GbE management switch. 29 Elastic Cloud Storage (ECS) H14071 1.10

Figure 24 - Example of 1 GbE Cisco Switch (Front View) 3.7.1.6 Disk Array Enclosure (DAE) For the U-Series, the DAEs are 4U enclosures that can have a maximum of 60 disks. Each DAE is connected to one node by one (Gen-1) or two (Gen-2) 6 GB/s SAS connection. For instance, there are four DAEs in a 4- node rack, and eight DAEs in an 8-node ECS rack. Currently, the supported drives are 6TB 7200 RPM SATA disks for Gen-1 and 8TB 7200 RPM SAS disks for Gen-2. Each DAE for the U-Series includes a hotswappable LCC. The LCC s main function is to be a SAS expander and provide enclosure services for all drive slots. The LCC independently monitors the environmental status of the entire enclosure and communicates the status to the ECS Element Manager. Note that the LCC is a hot-swappable component, but it cannot be replaced non-disruptively. The DAE has just one LCC, therefore, the disks in the DAE need to be brought offline before the LCC is replaced. For the D-Series models, the DAEs are 4U enclosures that have a maximum of 98 disks. It connects to a node by two12 Gb/s SAS connection. It supports 8TB or 10 TB 7200 RPM SAS disks. Entire chassis is cold serviceable which means replacing the drives or I/O modules would require temporarily shutting down the entire DAE and placing the node in maintenance mode. 3.7.1.7 ECS Node Connectivity There are up to four nodes in a 2U blade server, and each node has one or two SAS connections to a DAE (depending on the hardware generation used) and network interface cards (NIC) to each of the two 10 GbE data switches and to the 1 GbE management switch. Each node has two 10 GbE ports, which appear to the outside world as one port via NIC bonding. The 10GbE data switches in each rack will be connected to a customer-provided switch or backplane. Thus the data traffic will flow through the 10 GbE network. These public ports on the ECS nodes get their IP addresses from the customer s network, either statically or via a DHCP server. Customer applications connect to ECS by using the 10 GbE public IP addresses of the ECS nodes. The 1 GbE management port in a node connects to an appropriate port in the 1 GbE management switch and has a private address of 192.168.219.X. Each node also has a connection between its RMM port and a port in the 1 GbE switch, which in turn has access to a customer s network to provide out-of-band management of the nodes. To enable access for the RMM ports to the customer s network, ports 51 and/or 52 in the management switch are linked to the customer s network directly. The RMM port is used by Dell EMC field service personnel for monitoring, troubleshooting and installation. You can expand an ECS rack by linking one or more ECS racks to an existing rack also via ports 51 and 52 of the management switch. The 1 GbE switches in the racks are used for serially linking the racks. For more detailed information on connectivity, please refer to the ECS Hardware and Cabling Guide and other references stated earlier in this section. Figures 25 and 26, which depict the network cabling, as well as the figures in the hardware section also appear in the ECS Hardware and Cabling Guide. Also available for reference is the ECS Networking and Best Practices whitepaper. 30 Elastic Cloud Storage (ECS) H14071 1.10

Figure 25 Example of 10 GbE Network Cabling for 4 nodes Figure 26 Example of 1 GbE Network Cabling for 4 nodes 3.7.1.8 Upgrading ECS You can upgrade ECS by adding extra disks to existing DAEs or adding nodes and DAEs. Depending on model, each DAE can store maximum 60 or 98 disk drives, but disks are upgraded 5, 10, 15 or 28 per node at a time depending on model. Please refer to the ECS Hardware and Cabling Guide for more information on 31 Elastic Cloud Storage (ECS) H14071 1.10

the upgrade paths for each series of ECS. Also, check with Dell EMC s sales tools or representative to verify what upgrades are possible for a given configuration. 3.7.2 ECS Software on Industry Standard Hardware ECS software version 2.2.1 or later bundled with SLES 12 SP1 or later, Docker, JVM and tools can be installed on Dell EMC-approved 3rd-party industry standard hardware. There are two deployment offerings available: Certified - ECS software running on certified hardware. Custom ECS software running on hardware, operating system and tools outside of certified matrix The certified offering is targeted for customers needing small to large petabytes of object storage whereas custom is for big customer deals requiring a significant amount in petabytes of object storage. The value of deploying ECS software on commercial off the shelf hardware is that it reduces capital expense, prevents vendor lock-in and enables customers to build homogenous datacenter infrastructure with unified commodity hardware. Table 3 describes minimum general requirements to use ECS software solution on industry standard hardware. Table 2 ECS Software Minimum Requirements Type Requirement ECS Software Bundle Version 2.2.1 HF 1 or later bundled with SLES 12 SP1 or later, Docker, JVM and related tools. Server Nodes 5 nodes minimum Single 8-core minimum CPU per node 64 GB memory per node 10 disks per node (nearline HDD) Operating System Disk Networking Minimum 400 GB SSD** 10 GbE data (dual 10 GbE recommended) 1 Gbe management port (required) Disk to Node Connectivity HBA or HBA pass-thru mode recommended No RAID controller or a RAID controller that supports JBOD mode. All disks need to be in JBOD mode. Monitoring Integrate with Dell EMC ESRS for remote debugging/monitoring **NOTE: For DSS 7000 only, the minimum is 120GB SSD due to current availability. 32 Elastic Cloud Storage (ECS) H14071 1.10

The same data and management Arista switches used in the ECS appliance are used with the industry standard hardware. Customers are required to have engineering resources to install and manage the operating system and hardware. Current hardware that has been certified with ECS Software Bundle is defined in Table 4. Reference Architecture papers are also available for some of the certified hardware and can be downloaded by specified links in the table. Table 3 - Industry Standard Hardware Certified Industry Standard Hardware Certified HP Proliant SL4540 Gen 8 Dell R730XD Dell DSS 7000 In order to setup ECS on certified hardware, use the Precheck Guide to verify your environment meets minimum requirements. Then follow the standard ECS Appliance installation documentation to install the software. The Precheck Guide and the Installation Guides are available through SolVe desktop. Customers who would like to deploy ECS software on industry standard hardware can request for a quote via ASD Helpdesk. Custom configurations would require approval from product management and a qualification process. 3.7.3 Network Separation In ECS 2.2.1 and later, there is support for segregating different types of network traffic for security and performance isolation. The main types of traffic that can be separated include: Management Traffic Geo Replication Traffic Data Traffic There will be a new mode of operation called the network separation mode. When enabled during deployment, each node can be configured at the operating system level with up to three IP addresses or logical networks for each of the different types of traffic. This feature has been designed to provide the flexibility of either creating three separate logical networks for management, replication and data or combining them to either create two logical networks, for instance management and replication traffic is in one logical network and data traffic in another logical network. In ECS 3.1, if desired, a second logical data network for 33 Elastic Cloud Storage (ECS) H14071 1.10

CAS only traffic can be configured, allowing separation of CAS traffic from other types of data traffic like S3, etc. ECS implementation of network separation requires each logical network traffic to be associated with particular services and ports. For instance, the ECS portal services communicate via ports 80 or 443, so these particular ports and services will be tied to the management logical network. As of ECS 3.1, a second data network can be configured; however, it is for CAS traffic only. The table below highlights the services fixed to a particular type of logical network. For a complete list of services associated with ports, refer to the ECS Security Configuration Guide. Table 4 - Services associated with logical network. Services ECS Portal, Provisioning, metering and management API and ssh Data across NFS, Object and HDFS Replication data, XOR ESRS (Dell EMC Secure Remote Services) DNS, NTP, AD, SMTP Logical Network Management Network Data network 2 nd Data network (CAS traffic only) Replication Network Based on the network that the ESRS Gateway is attached (data or management network) Management network. Network separation is achievable via logical using different IP addresses, virtual using different VLANs or physical using different cables. The command setrackinfo is used to configure the IP addresses and VLANs. Any switch-level or client-side VLAN configuration is the customer s responsibility. For physical network separation, customers would need to submit a Request for Product Qualification (RPQ) by contacting Dell EMC Global Business Service (GBS). For more information on network separation refer to the following documents: - ECS Networking and Best Practices provides a high level view of network separation. - ECS Networking and Node IP Change document in SolVe Desktop (Procedure Generator) for configuration and setup instructions. As of ECS 3.1, in addition to new installations, network separation is available for existing installations. Support Depending on whether you run ECS software on an ECS appliance or industry standard hardware, the support models differ. Table 5 describes the support model offered for each type of hardware deployment. Table 5 - Software and Hardware Support Type ECS Appliance Industry Standard Hardware (Certified) Industry Standard Hardware (Custom) Software Package ECS Software ECS Software v2.2.1 ECS Software v2.2.1 Hotfix 1 or later Hotfix 1 or later Operating System SLES 12 SP1 or SLES 12 SP1 or later Customer 34 Elastic Cloud Storage (ECS) H14071 1.10

Hardware Infrastructure Network Infrastructure later provided by Dell EMC Dell EMC provided hardware rack Dell EMC provides fully managed rack Bundled with ECS Software Dell EMC certified and approved hardware Uses the same network infrastructure as in Dell EMC ECS appliance Customer Customer Hardware Maintenance Dell EMC Customer Customer Hardware Issue Resolution Dell EMC Customer Customer Operating System Patch Deployment and Management Dell EMC Dell EMC Customer Operating System Issue Resolution Dell EMC Dell EMC Customer ECS Software Installation and Upgrade Using ECS Tools Dell EMC Dell EMC Dell EMC ECS Storage Engine and Access Protocol Issue Resolution Dell EMC Dell EMC Dell EMC Service Reliability and Availability Issue Resolution Dell EMC Dell EMC Dell EMC Assistance With Other Storage Issues (e.g. cluster expansion, disk replacement, etc.) Dell EMC Dell EMC Dell EMC 35 Elastic Cloud Storage (ECS) H14071 1.10

4 Security ECS security is implemented at the administration, transport and data levels. User and administration authentication is achieved via Active Directory, LDAP methods, Keystone, or within the ECS portal. Data level security is done via HTTPS for data in motion and/or server-side encryption for data at rest. 4.1 Authentication ECS supports Active Directory, LDAP and Keystone authentication methods to provide access to manage and configure ECS; however, limitations exist as shown in Table 6. For more information on security, see the ECS Security Configuration Guide. Table 6 - Authentication Methods Authentication Method Active Directory LDAP Keystone Supported AD Groups are supported for management users Supported for Object Users (Self-service keys via API) Multi-domain is supported. LDAP Groups are not supported for management users LDAP is supported for Object Users (Self-service keys via API) Multi-domain is supported. RBAC policies not yet supported. No support for un-scoped tokens No support of multiple Keystone Servers per ECS system 4.2 Data Services Authentication Object access using RESTful APIs is secured over HTTPS (TLS v1.2) via specific ports, depending on the protocol. All incoming requests are authenticated using defined methods such as Hash-based Message Authentication Code (HMAC), Kerberos, or token authentication methods. Table 7 below presents the different methods used for each protocol. Table 7 - Data Services Authentication Protocols S3 Object Swift Atmos CAS HDFS NFS Authentication Methods V2 (HMAC-SHA1), V4 (HMAC-SHA256) Token Keystone v2 and v3 (scoped, UUID, PKI tokens), SWAuth v1 HMAC-SHA1 Secret Key PEA file Kerberos Kerberos, AUTH_SYS 36 Elastic Cloud Storage (ECS) H14071 1.10

4.3 Data At-Rest-Encryption (D@RE) ECS version 2.2 and later supports server-side encryption. Server-side encryption follows the FIPS-140-2 Level 1 compliance, AES256 algorithm. Key features of ECS Data-at rest-encryption include: Low touch encryption at rest enabled easily through the ECS Portal Namespace- and bucket-level control with transitivity encryption can occur at namespace or bucket level Automated key management S3 encryption semantics support constructs can be used to allow object encryption, e.g. x-amzserver-side-encryption A valid license is required to enable server-side encryption via the ECS Portal or through the ECS REST API. Each namespace, bucket and object has an associated key that is auto-generated at the time of creation. Only data and user-defined metadata will be encrypted inline prior to being stored on disks. 4.3.1 Key Management Encryption can be enabled at the namespace, bucket and/or object levels. There are two types of keys: User provided keys via S3 Header key is provided by user through the header in the S3 API System-generated keys randomly generated and hierarchically structured For S3 users and applications using the S3 REST APIs, object-level encryption can be done by either systemgenerated or user-specified keys. If the S3 encryption header provides the key, then encryption of object is done using the user-provided key. ECS validates the key provided for updates, appends, and reads is the same as the key used for object creation. If the encryption header is provided without a key, then the systemgenerated keys are used to encrypt the data. System-generated keys at each level are auto-generated and encrypted using the key of its immediate parent. The master key is encrypted by utilizing asymmetric public-private encryption. The public and private keys are used to encrypt and decrypt the master key. Keys are generated and encrypted in a hierarchical order: Public-Private Key pair a pair of public-private keys generated for each ECS system. The private key is stored in the system root disk of the node and the public key is stored with the master key on the ECS commodity disks. Master Key randomly generated key encrypted using the public key. Namespace Key randomly generated namespace key encrypted using the master key. Bucket Key randomly generated bucket key encrypted using the namespace key. Object Key randomly generated object key encrypted using the bucket key and object id. The private key for decryption of master key is stored on system disks and the other encrypted keys are stored in logical tables and in chunks, similar to data. They are also triple-mirrored, like data. When an object read or write request comes in, the node servicing the request traverses the key hierarchy to encrypt or decrypt the object. It uses the private key that s common to all nodes to decrypt the master key. Figure 27 provides a pictorial view of the key hierarchy. 37 Elastic Cloud Storage (ECS) H14071 1.10

Figure 27 - Data Encryption using System Generated Keys In a geo-replicated environment, when a new ECS system joins an existing system (referred to as a federation), the master key is extracted using the public-private key of the existing system and encrypted using the new public-private key pair generated from the new system that joined the federation. From this point on, the master key is global and known to both systems within the federation. In Figure 28, the ECS system labeled VDC 2 joins the federation and the master key of VDC 1 (the existing system) is extracted and passed to VDC 2 for encryption with the public-private key randomly generated by VDC 2. Figure 28 - Encryption of Master Key in Geo Replicated Environment. 38 Elastic Cloud Storage (ECS) H14071 1.10

5 Data Integrity and Protection The most common approach for data protection within storage systems is RAID. ECS, however, does not utilize RAID; it employs a hybrid of triple mirroring of data, metadata, and index and also erasure coding of data for enhanced data protection and reduction of storage overhead. For data integrity, ECS utilizes checksums. 5.1 Triple-mirrored All types of information relating to objects, such as data, metadata, and index (B+ tree and journal logs) are written to chunks. At ingest, the chunks are triple mirrored to three different nodes within the ECS system as shown in Figure 29. This technique of triple mirroring allows for data protection of the data in case of a node or disk failure. For data chunks, one of the triple mirror copies is erasure coded in place for performance. In place erasure coding means that one copy is written in fragments that are spread across different nodes and disks. This optimizes performance by not requiring data to be redistributed after the chunk is sealed and parity is calculated. Figure 29 - Triple Mirroring of Chunks 5.2 Erasure Coding Erasure coding provides enhanced data protection from a disk or node failure in a storage efficient fashion compared with conventional protection schemes. The ECS storage engine implements the Reed Solomon 12+4 erasure-coding scheme, in which a chunk is broken into 12 data fragments and 4 coding (parity) fragments. The resulting 16 fragments are dispersed across nodes at the local site. The data and coding fragments of each chunk are equally distributed across nodes in the cluster. For example, with 8 nodes, each node has 2 fragments (out of 16 total). The storage engine can reconstruct a chunk from any 12 of the 16 fragments. ECS 2.2 and later includes a cold archive option, for which a 10+2 erasure coding scheme is used. In this protection scheme, each chunk is broken into 10 data fragments and 2 coding (parity) fragments, which allows for better storage efficiencies for particular use cases. All data in ECS is erasure coded except the index and system metadata. The index provides location to objects and chunks and is frequently accessed; hence, it is always kept in triple-mirrored chunks for protection. ECS requires a minimum of four nodes to be able to conduct the default erasure coding and six nodes for the cold archive option, in which a 10+2 scheme is used instead of 12+4. Erasure coding stops when the number of nodes goes below the minimum for the EC scheme. When a chunk is full (128MB) or after a set period of 39 Elastic Cloud Storage (ECS) H14071 1.10

time, it s sealed, parity is calculated and the coding (parity) fragments are written to disks across different storage nodes. Erasure coding is conducted as a background process. After erasure coding completes, the mirrored copies are discarded and a single erasure coded copy persists. In the erasure-coded copy, the original chunk data remains as a single copy that consists of 16 fragments (12 data, 4 code) dispersed throughout the cluster. For small objects, ECS doesn t need to read all the fragments of the data -- it can directly read the object from the data fragment which contains it. ECS only uses the code fragments for chunk reconstruction when a failure occurs. For objects greater than 128MB in size, erasure coding is done in-line and erasure-coded data is directly written to the disk. This is done to enhance performance and decrease disk usage. Figure 30 gives a view of how a triple mirrored chunk is replaced by an erasure coded chunk. Figure 30 Example of Erasure Coding Layout for 12+4 Scheme In ECS 2.2 and later, when additional nodes are added into a configuration to upgrade capacity, ECS redistributes erasure coded chunks in a background task. This will have minimal impact on system performance; however, there will be an increase in inter-node traffic during rebalancing. Balancing of the logical tables onto the new nodes added is also conducted. Newly created journal and B+ tree chunks are evenly allocated on old and new nodes going forward. Redistribution enhances local protection by leveraging all of the resources within the infrastructure. NOTE: Adding new disks to existing nodes does not trigger data rebalancing. As a best practice, it is recommended not to wait until the storage platform is completely full before adding drives or nodes. A reasonable storage utilization threshold is 70% taking consideration daily ingest rate and expected order, delivery and integration time of added drives/nodes 5.3 Checksums Another mechanism ECS uses to ensure data integrity is to store the checksum for data written. Checksums are done per write-unit, up to 2 MB. So, checksums can occur for one object fragment for large object writes or on a per-object basis for small object writes of less than 2MB. During write operations, the checksum is calculated in memory and then written to disk. On reads, data is read along with the checksum, and then the checksum is calculated in memory from the data read and compared with the checksum stored in disk to determine data integrity. Also, the storage engine runs a consistency checker periodically in the background and does checksum verification over the entire data set. 40 Elastic Cloud Storage (ECS) H14071 1.10

5.4 Compliance To meet corporate and industry compliance requirements (SEC Rule 17a-4f) for storage of data, ECS implemented the following components: Platform hardening addresses security vulnerabilities in ECS such as platform lockdown to disable access to nodes or cluster, all non-essential ports (e.g. ftpd, sshd) are closed, full audit logging for sudo commands and support for ESRS(Dell EMC Secure Remote services) to shut down all remote access to nodes. Compliance Reporting a system agent that reports system s compliance status such as Good indicating compliance or Bad indicating non-compliance. Policy Based record retention and rules limiting the ability to change records or data under retention using retention policies, time-period and rules. Advanced Retention Management (ARM) to meet Centera compliance requirements additional retention rules were defined for CAS only. Event Based Retention enables application to specify retention period that will start when a specified event occurs Litigation Hold enables application to temporarily prevent deletion of an object that is subject to an investigation or legal action. Min/Max Governor enables system administrator to specify a minimum and maximum value for the default retention period on a per bucket basis. Compliance is enabled at the namespace level and retention periods at the bucket level. Since compliance requirements certify the platform, the compliance feature is available for ECS Appliance only. For information on enabling and configuring of compliance in ECS please refer to the Data Access Guide and ECS System Administrator s Guide. 41 Elastic Cloud Storage (ECS) H14071 1.10

6 Deployment ECS can be deployed as a single site or in a multi-site configuration. The building blocks of an ECS deployment include: Virtual Data Center (VDC) a geographical location defined as a single ECS deployment within a site. Multiple VDCs can be managed as a unit. Storage Pool a storage pool can be thought of as a subset of nodes and its associated storage belonging to a VDC. An ECS node can belong to only one storage pool; a storage pool can have any number of nodes, the minimum being four. A storage pool can be used as a tool for physically separating data belonging to different applications. Replication Group defines where storage pool content is protected and locations from which data can be read or written. Local replication groups protect objects within the same VDC against disk or node failures. Global replication groups span multiple VDCs and protect objects against disk, node, and site failures. Namespace - a namespace, which is conceptually the same as a tenant, is a logical construct. The key characteristic of a namespace is that users from one namespace cannot access objects belonging to another namespace. Namespaces can represent a department within an organization or a group within a department. Buckets container for object data. Buckets are created in a namespace to give applications access to data stored within ECS. In S3, these containers are called buckets and this term has been adopted by ECS. In Atmos, the equivalent of a bucket is a subtenant ; in Swift, the equivalent of a bucket is a container, and for CAS, a bucket is a CAS pool. Buckets are global resources in ECS. Where the replication group spans multiple sites, a bucket is similarly replicated across sites. In order to be able to deploy ECS, certain infrastructure requirements need to be reachable by the ECS system. Authentication Providers users (system admin, namespace admin and object users) can be authenticated using Active Directory or LDAP or Keystone DNS Server Domain Name server or forwarder NTP Server Network Time Protocol server. Please refer to the NTP best practices for guidance on optimum configuration SMTP Server (optional) Simple Mail Transfer Protocol Server is used for sending alerts and reporting from the ECS rack. DHCP server only if assigning IP addresses via DHCP Load Balancer - (optional but highly recommended) evenly distributes loads across all data services nodes. Load balancers can use simple algorithms such as random choice or round robin. More sophisticated load balancers may take additional factors into account, such as a server's reported load, response times, up/down status, number of active connections, geographic location and so on. The customer is responsible for implementing load balancers; customers have several options including Manual IP allocation, DNS Round Robin, Client-Side Load Balancing, Load Balancer Appliances, and Geographic Load Balancers. The following are brief descriptions of each of those methods: 42 Elastic Cloud Storage (ECS) H14071 1.10

- Manual IP Allocation - Data node IP addresses are manually distributed to applications. This is not recommended because it does not evenly distribute loads between the nodes and does not provide any fault-tolerance if a node fails. - DNS Round-Robin - With DNS Round-Robin, a DNS name is created for ECS and includes all of the IP addresses for the data nodes. The DNS server will randomly return the IP addresses when queried and provide some pseudo-load balancing. This generally does not provide fault-tolerance because you would need to remove the IP addresses from DNS to keep them out of rotation. Even after removing them, there is generally some TTL (time-tolive) issues where there is a delay to propagate the removal. Also, some operating systems like Windows will cache DNS lookups and can cause "stickiness," where a client keeps binding to the same IP address, reducing the amount of load distribution to the data nodes. - Physical or Virtual load balancing- This option is the most common approach to load balancing. In this mode, an appliance (hardware or software) receives the HTTP request and forwards it on to the data nodes. The appliance keeps track of the state of all of the data nodes (up/down, # of connections) and can intelligently distribute load amongst the nodes. Generally, the appliance will proactively "health check" the node (e.g. GET/?ping on the S3 head) to ensure the node is up and available. If the node becomes unavailable it will immediately be removed from rotation until it passes a health check. Another advantage to this kind of load balancing is SSL termination. You can install the SSL certificate on the load balancer and have the load balancer handle the SSL negotiation. The connection between the load balancer and the data node is then unencrypted. This reduces the load on the data nodes because they do not have to handle the CPU-intensive task of SSL negotiation. - Geographic load balancing - Geographic load balancing takes Physical or Virtual Load Balancing one step further: it adds load balancing into the DNS infrastructure. When the DNS lookups occur, they are routed via an "NS" record in DNS to delegate the lookups to a load balancing appliance like the Riverbed SteelApp. The load balancer can then use Geo-IP or some other mechanism to determine which site to route the client to. If a site is detected to be down, the site can be removed quickly from DNS and traffic will be routed to surviving sites. For examples on how to deploy ECS with load balancers, refer to the following whitepapers: - ECS with HAProxy Load Balancer Deployment Reference Guide - ECS with NGINX (OpenResty) Deployment Reference Guide 6.1 Single-site Deployment In a single site, storage pools are defined, and then the Virtual Data Center (VDC) is created with namespaces and buckets. Figure 31 shows one storage pool in a VDC with a namespace containing two buckets. 43 Elastic Cloud Storage (ECS) H14071 1.10

Figure 31 Single Site Deployment Example 6.2 Multi-site Deployment In a multisite deployment, more than one VDC is managed as a federation and/or geo-replicated. In a georeplicated deployment, data can be read or written from any active site within the defined replication group. This section describes this type of deployment. 6.2.1 Geo-Federation In essence, geo-federation means managing a geographically distributed environment as a single logical resource. Inside ECS, the term refers to the ability to federate multiple sites or VDCs. The obvious benefits are ease of management and the ability to use resources from multiple data centers. To manage or monitor a single site, the administrator simply logs into one of the nodes in the VDC. This is equivalent to having a single-site federation. The administrator can subsequently add other VDCs to the federation by navigating to the VDC in the ECS portal at the first VDC. Further, customers can choose how data is replicated across sites for enhanced data protection. To perform some administrative tasks such as creating storage pools or viewing performance statistics, one has to log into each VDC individually. 6.2.2 Geo-Replication As discussed earlier, ECS offers erasure coding to help provide enhanced data durability without the overhead of storing multiple copies of the data. However, this does not protect against site failures/outages. Geo-replication provides enhanced protection against site failures by having multiple copies of the data, i.e., a primary copy of the data at the original site and a secondary copy of the data at a remote site/vdc. Both the primary and replicated copies of data at other sites are individually protected via erasure coding or triplemirrored chunks. This means each copy has protection from local failures, such as disk or node failure. Replication is done by site owning object and is an asynchronous process. Each site is responsible for local data protection. Replicated data is first encrypted (AES256) and then sent to another site via HTTP. By 44 Elastic Cloud Storage (ECS) H14071 1.10

default, data is replicated to one other site within the replication group, however if the Replicate to All Sites option is enabled in three or more sites, then the data is replicated to all sites. Geo-replication ensures that data is protected against site failures/disasters. Unlike other solutions in the marketplace, ECS does not generate WAN traffic while recovering from local disk failures. ECS gives customers the option to link geographically dispersed systems and replicate data among these sites across the WAN. Several strategies, such as geo-caching as well as accessing the physically closest array, reduce WAN traffic for data access. A replication (RG) is defined in a geo-replicated deployment and defines where data is protected and where it can be accessed from. ECS supports both local and global replication groups. A local replication group contains a single VDC and protects data within the same VDC against disk or node failures. Global replication groups contain more than one VDC, and protect data against disk, node and site failures. Replication groups are assigned at the bucket level. Figure 32 shows an example of a three site configuration with different replication group policies assigned to different buckets. Figure 32 - Local and Global Replication Groups Example In this example clients accessing: VDC1 have access to all buckets. VDC2 and VDC 3 have access only to buckets 01 and 02. ECS supports geo-active replication and starting in ECS version 3.1, geo-passive replication. 6.2.2.1 Geo-Active Replication One key feature of ECS s geo active replication is the ability to read and write data from any VDC within a replication group. In ECS, data is replicated asynchronously to multiple VDCs within a replication group configured with geo active replication. The challenge this poses is consistency of data across sites or VDCs. 45 Elastic Cloud Storage (ECS) H14071 1.10

ECS ensures strong consistency by fetching the metadata from the VDC that created the object data. Thus, if an object is created in site 1 and is read from site 2, ECS checks with site 1, who is the object owner, and validates that the copy replicated at site 2 is the latest version of the data. If not, it fetches the data from site 1; otherwise, it uses the data from site 2. The data flow of reads in a geo-replicated environment illustrated in Figure 33 below is as follows: 1. Site 2 does a read request. In this example, Site 1 is the object owner. 2. Read would require checking metadata from object owner to validate if Site 2 has the latest copy of data. 3. If not, then Site 1 will send the data to Site 2. 4. Data sent to client. Figure 33 - Read Data Flow in Geo Replicated Environment Example The data flow of writes in a geo-replicated environment in which two sites are updating the same object is shown in Figure 34. In this example, Site 1 who initially created the object becomes object owner. The data is mirrored or erasure coded and the journal is written as usual to the local site. The data flow for writes include: 1. Site 2 updates the same file. Site 2, first writes the data locally (the data is mirrored or erasure coded locally) 2. Site 2 synchronously updates the metadata (journal write) with the object owner, Site 1 and waits for acknowledgement of metadata update from site 1. 3. Site 1 acknowledges the metadata write to Site 2. 4. Site 2 acknowledges the write to the client. 46 Elastic Cloud Storage (ECS) H14071 1.10

Note: Site 2 will asynchronously replicate the chunk to site 1 (owning site) as usual. If the data is not yet replicated to the owning site and the owning site wants to read the data, then it will get it from the remote site Figure 34 - Update of Same Object Data Flow in Geo-replicated Environment Example In both read and write scenarios in a geo-replicated environment, there is latency involved in reading and updating the metadata from object owner and retrieving data from the object owner during a request. 6.2.2.2 Geo-Passive Replication Starting in ECS 3.1 a replication group can be configured as geo-passive which designates two source sites and a third site to be used as a replication target only. This third site is used for recovery purposes only, it does not allow for direct client access. The benefits of geo-passive replication include: It can optimize storage efficiency by increasing the chances of XOR operations running by ensuring writes from both source sites go to the same replication target. It allows the administrator to control where the replication copy of data exists, such as in a backup to cloud scenario. Figure 35 shows an example of a geo-passive configuration whereby VDC 1 and VDC 2 are source sites and both are replicating their chunks to the replication target, VDC 3. 47 Elastic Cloud Storage (ECS) H14071 1.10

Figure 35 - Example of Client Access and Replication Paths for Geo-Passive 6.2.3 Geo-Caching ECS optimizes response times for accessing data from a remote site by predesignating a percentage of disk space to cache objects that do not exist locally. This is important for customers with multi-site access patterns, specifically a replication group composed of three or more sites where data is often fetched from a remote site. Consider a geo-replicated environment with Sites 1, 2 and 3. An object, Object A, is written to Site 1 and the secondary copy of the object resides at Site 2. In this scenario, a read for the object at Site 3 needs to fetch the object data from either Site 1 or Site 2 to honor the read. This leads to elongated response times, especially in environments with multi-site access patterns. With geo-caching frequently accessed objects would see a reduced response time after the initial copy of the object is cached locally. The cache implements a Least Recently Used (LRU) algorithm and cache size is adjusted when nodes/disks are added to the storage pool. 6.2.4 Temporary Site Outage Temporary site failure refers to either a failure of the WAN connection between two sites or a failure of an entire site (such as a natural disaster or power failure). ECS can detect and automatically handle any such temporary site failures. VDCs in a geo-replicated environment establish a heartbeat mechanism. Sustained loss of heartbeats for a preset duration is indicative of a network outage and the system adjusts its behavior accordingly. ECS provides options that affect how objects can be accessed during a temporary site outage (TSO). These options include: Default retain strong consistency by allowing access to data owned by sites that are accessible and prevents access to data owned by an inaccessible site. 48 Elastic Cloud Storage (ECS) H14071 1.10